Skip to content

Basics

AI automation in SMEs: where it pays off and where it does not

By Dominik Natter · Last updated on 5 November 2026 · 2 min read

In short

AI automation pays off in high-volume processes where people read unstructured content and transfer it into a system: invoices, enquiries, emails, forms. It does not pay off when the volume is small or when errors cannot be caught by a check. What matters is human approval for uncertain cases, an audit trail and a written agreement on where the data is processed.

What language models can do reliably today

  • Read documents: invoices, orders, delivery notes, forms.
  • Classify messages: type of request, urgency, language, team responsible.
  • Match content: line items in an enquiry to items in the master data.
  • Write drafts: replies, quote texts, summaries.
Starting price for email inbox triage, up to CHF 30,000
CHF 12,000
Starting price for invoice capture with bank reconciliation, up to CHF 60,000
CHF 25,000
Weeks to implement, depending on the use case
4 to 14

Typical use cases and effort

Use casePrice rangeDuration
Email inbox triageCHF 12,000 to 30,0004 to 8 weeks
Creating quotes from enquiriesCHF 20,000 to 50,0006 to 12 weeks
Invoice capture with bank reconciliationCHF 25,000 to 60,0008 to 14 weeks

Where caution is called for

Language models do not calculate reliably and do not know your prices. Anything that has to be exact belongs in your systems: amounts, terms, stock levels. The AI reads and suggests, the ERP calculates and posts.

How AI, people and the ERP share the work

  1. 01

    Content comes in

    Invoices, enquiries, emails and forms arrive, mostly unstructured and in large numbers.

  2. 02

    AI reads and suggests

    The language model reads documents, classifies messages and matches line items to items in the master data.

  3. 03

    A person approves unsure cases

    Uncertain cases go to a person for approval instead of running through automatically.

  4. 04

    ERP calculates and posts

    Amounts, terms and stock levels come from your systems. The ERP calculates and posts.

  5. 05

    Audit trail records everything

    An audit trail shows what the AI suggested and who approved what.

Workflow tool or business application?

Tools such as n8n connect systems quickly and at low cost. For a manageable workflow that is the right choice. As soon as permissions, an audit trail, high volumes, error handling or a user interface for approvals are needed, the workflow becomes an application that has to be developed, tested and operated.

A workflow tool reaches its limits if

  • Permissions have to be managed.
  • An audit trail is required.
  • Volumes are high.
  • Errors have to be handled properly.
  • People approve cases through a dedicated user interface.

A workflow tool is enough if

  • The workflow is manageable.
  • Systems need to be connected quickly and at low cost.
  • An error is cheap or certain to be noticed.

The difference shows in day-to-day operation: what happens when the ERP does not respond, an invoice arrives twice or the model is unsure?

Data protection in five points

  • Agree processing in a European region.
  • Rule out the use of your data for training by contract.
  • Give the model only the data the task requires.
  • Define how long logs are kept and when they are deleted.
  • Train staff and document the human approval step.

Legal framework

The GDPR applies in Liechtenstein and the EU, the revised Federal Act on Data Protection in Switzerland. Both require transparency about which data is processed and for what purpose. The EU AI Act additionally regulates the use of AI by risk level.

Where this information comes from

Price ranges and duration

The price ranges and time frames are estimates for the three typical use cases. A price only becomes binding with a discovery phase.

Legal framework

The information refers to the GDPR, the revised Swiss Federal Act on Data Protection and the EU AI Act. It gives an overview and does not replace a case-by-case assessment.

As of: November 2026.

Frequently asked questions

How do you stop the AI from making mistakes?

By never letting it decide alone where mistakes are expensive. Every extraction is given a confidence score, uncertain cases go to a person for approval, and every decision is logged. Before go-live we measure the accuracy rate on your real data.

Where does our data go when AI is involved?

We set that out in writing for each project. The default is processing in a European region, under contracts that rule out any use of your data for training. For sensitive data, self-hosted models are an option. Hosting and database are located in the EEA, and processing is designed around the GDPR and the Swiss Federal Act on Data Protection.

Do you work with n8n or build your own software?

Both, depending on the case. n8n suits manageable workflows between existing systems, and your team can help maintain it. As soon as audit trails, permissions, high volumes or a dedicated user interface are needed, a purpose-built application is more robust. We often combine the two.

Matching services:AI automation

Let us talk about your case.

20 minutes are enough to find out whether a custom solution is worth it for you.